Abstract
BACKGROUND: Rotator cuff injury is a common cause of shoulder pain. Precise and efficient measurement of morphological parameters is necessary in the clinical diagnosis and evaluation of shoulder disorders. However, manual measurement is a time-consuming and labor-intensive task, with low inter-observer reliability. The automatic measurement of radiographic parameters in supraspinatus outlet radiographs has not been reported yet. Thus, the objective of this study was to use a cascaded High-Resolution Net (HRNet) model based on deep learning (DL) algorithms to automatically measure morphological parameters from supraspinatus outlet radiographs and assess its performance. It was intended for use in early screening of patients with rotator cuff disease and to guide them to further consultation. METHODS: This cross-sectional study collected 1,668 supraspinatus outlet radiographs from the picture archiving and communication system of Gansu Provincial Hospital of Traditional Chinese Medicine and the Affiliated Hospital of Gansu University of Chinese Medicine. Among them, 521 images were provided for test datasets and 1,147 images were provided for a model training dataset and validation dataset. Landmarks were annotated for acromio-humeral interval (AHI), acromial tilt (AT), and 3 lines in Park's acromial classification (line huo-acrf, line acro-acro1, and line huo-acro1). R4 radiologist reviewed the means of 3 radiologists as a reference standard. Model performance was assessed by calculating the percentage of correct key points (PCK), intra-class correlation coefficients (ICCs), Pearson's correlation coefficients, mean absolute error, and root mean square error. The reliability of R1, R2, R3, AI with R4 and inter-observer reliability of R1, R2, and R3 for acromial morphology classification were assessed by Cohen's kappa coefficient. RESULTS: Within the 3-mm threshold, the PCK of the model ranged from 74% to 100%. Compared to the reference standard, the model had reliable measurement of AHI, AT, line huo-acrf, line acro-acro1, line huo-acro1 (ICC =0.73-0.94) and moderate reliability of acromial morphology classification (k=0.50-0.56). CONCLUSIONS: The cascaded HRNet developed in this study can automatically measure morphological parameters of the shoulder. It may aid early clinical screening for shoulder disorders and assist physicians in treatment decisions.